PERCEPTION AND PRODUCTION OF TWO-STOP-CONSONANT SEQUENCES

Bruno H. Repp

Abstract. The duration of the silent closure interval required toperceive two stop consonants in a VC1C2V sequence depends, to someextent, on their places of articulation. In production, too, theduration of the closure interval varies systematically with place.Hqwever, there appears to be little relation between the patterns ofvariability in production and in perception. Moreover, two analo­gous perceptual experiments--one using synthetic stimuli, the other,natural speech--yield quite different results. Thus, variations inthe amount of closure required to perceive two successive stops seemto be governed by stimulus-specific acoustic factors, not by aninternal representation of articulatory patterns or constraints.This conclusion is further supported by the unexpected finding thatsome listeners do not require any closure for accurateperception of both stops.

INTRODUCTION

Lisker (1957) first reported that, when the waveforms of naturallyproduced Ir~1 (with Igl unreleased) and Ib%dl are abutted without anyintervening silence (which serves to indicate oral closure), listeners hearIr~b"J:d/--that is, they fail to the first ( ) stopconsonant. This effect was later rediscovered Abbs (1 and has, morerecently, been investigated in detail (Dorman, & Liber-man, 1979; Raphael & Dorman, in press; Repp, 1978, 1 1979b, 1980;Rudnicky & Cole, 1978). These studies used both and natural speech,and a variety of stop-consonant combinations and vocalic contexts 0 Severalstudies assessed what closure duration is needed between the VC1 andC2V waveforms to perceive both consonants on 50 of the trials; atypical value for this boundary on a continuum of ng silentclosure durations is msec However, the of the phenomenon isstill far from clear

Two ities may be di One is that the effect inquestion aud e.g., it be due to interference of thecues for the second (the formant transitions out of the with the

of the cues for the (the formant transitions into the

NICHD HDO 1 and BRSLaboratories. Most 0 the work was done my

thank Kuhn ssion to draw on hisLisker Vi

Grant toable assistant,collection of VCV

contributed

[HASKINS LABORATORIES on ( 1 )]

177

different stimuliof both

about relevant auditoryother is

appear's to beso, then

consonantsaverage (or, the

Neither ofin recent

closure)--cf0 Massaro (1975)0 If so, any variationsin the amount of closure necessary for accurateshould be reference to what is knownprocesses such as backward or gap detection0that mirrors articulation more or dithe case with many other phenomena invariations in the closure duration needed toshould be correlated with similar variations in the

closure duration in VCthese al ternatives has beenstudies of the influence of three( t duration, and itude ofof the ( 9

absence acoustic dataas to the the

tion

msec in

shorteracoustic observations

towards ato

sequences 0that to

178

less likelydepend on

articulatoryrements as a

work in thein thebecausein the

the9 and

ofoccurs

should be invariant across different vocalic contexts 0 In thecase that the articulatory dynamics of two-stop sequencesthe vocalic environment t the question becomes whether

correspond in any way to changingfunction of vocalic context If psychoacoustic factors

suppression of the first stop, considerable variabof results might be across different vocalic contexts

the acoustic properties of the stimuli rad withsurrounding vowels; in ar, the formant transitions

aces of of the two may in extent,direction to the auditory however, thevariabil observed in should relation to whatin

Thus, the stud address three issues: (1) Does theindeed vary across different combinations of as earlier

studies suggest and if so is this pattern of results stable across differentvocalic contexts? ) Do closure durations in naturalances vary across different combinations of thisstable across different vocalic contexts? any consistent

between the observed in in

ences between his that of of

179

fundamental40-140 msec

, andi tudes and

This ied, with ICil

itude

had afor

naturalAll

of 120 Hz.. All CV syllables were 290 msec long andcontour that at 120 Hz, remained

(depending upon the individual stimulus, as copied fromthen fell steadily to a value between 94 and 105 Hz ..formant ectories remained as traced from naturallower output amplitudes for ICul than for ICD/ and IdCI

i tudes in between.. ( ,1979b, showed that stimul usa minor role in the used here .. )

) ,to 115fferent

the Haskinsidentical

stimul i were d zed at 10Three test were then created,

(/0.1, Iii lui) whichsequence of the

in which each stimulof 3 sec.. The stimuli

AllLaboratories PCMfor vowel of the CVEach contained( , Io.d/, , IbVIwith interstimulus intervals (ISIs)of the test consisted of the sixwi th silent closure intervalsmsec.. The resul

zations, with

VC-CV combinations: Two-stop vs. one-stop The responses toVC-CV combinations were first scored in terms vs. one-stopresponses, regardless of whether the responses were correct (i.e., theequivalent of C1, C2, or C1C2) or not. (Exclusion of errors would havedistorted the data because of certain systematic misidentifications, which arediscussed below.) All VC-CV combinations showed the increase in two­stop responses as the silent interval increased in duration. The boundaryvalues (50-percent cross-over points) for all but two of the labelingfunctions fell between 55 and 80 msec. Two functions, however, stood out-­those for lagb~1 and ladbal; these stimuli required much less silence for both

to be heard, and they received a number of two-stopeven at the shortest sil ence duration. Note that both stimul i

contain back-to-front movements of of articulation, n wi thRaphael and Dorman (in

Figure , summarizes the data in terms of single-stopresponses, averaged across all silence durations--a measure that takes intoaccount differences in the lower and upper of the responsefunctions. (However, a plot in terms of boundary values a very similar

.) It can be seen that the deviant resul ts for Idbl and I in theICo.! set have no parallel in the ICil and leul sets; , they are

fic to the ICal stimuli (to Ibo./ in particular)" The thatfront-back sequences (the first three stimuli on the abscissa ,)would have lower boundary values (i"e., more single-stop back-front sequences (the last three stimuli on the absci inthe ICil and ICul sets, and partially supported in set, since

did not have a low value.

ficantand

ficant effect of ei therinteraction of these two.£ <. ,but difficul t to

for la and led tovariance" after

, there was novocalic context; theficant, 6,48)::: 3.0,

The deviant resul tseffects in an is ofI stimuli from theconsonant combinations orfactors was

VC-CVboundaries(/o.db~/,

IV,thedipossible

tended to havesets, even if the

somewhat shorterextreme cases

is

s

181

f-'coN

gd

6.

011111111111111110 lui

6. 6./010 ......0/ il

0,#..." ###o ###~ ##,

"" '"" '0""""",

\ ,";10

\. "o

e 1. Percent responses (averaged over all silence durations)to the combinations (synthetic speech) $

for by C2 confusions between Ibil and Idi/ @ Many of the C1responses could be predicted from the way the isolated stimulus componentswere ed, ex for small occurring in response to IQdbaland IQgb~@ Note that all these cases involve labial stops in secondposition; thus, labial formant transitions seemed to be lesseffective in wi th confl lable-final transitions than

and velar transitions@

the error responses (responses other than thecould be icted from the misidentifica-

There were certain unpredicted errors,included "bg" responses to

), which consti tuted thestimuli (total 9 ; and

, which made up about two thirds11@2 percent)@ These errors involve

in which the first stop waswas misidenti fied as "ab" in

a form of perceptual contrast

A ofents of C1, and

tions of the monosyl ichowever, that showed up withIe. and to (large of error responses to"bd" responses to , , andof the errors to these stimuli (total:al veolar-velar combinations (in eithermislabeled as "b" @ (Neither IQdl norisolation @) We may be deal here wi th(cf @ , 1978)

acoustic measurements of natural VC 1C2V utterances,is rel between the amount of

and the durations of closure periods inthere have been studies of closure durationsintervocalic of se-

(Note 1) @to

, whereas theNevertheless,

that total closure durations werewhen it was labial or velar;was velar than when it was

an effect of vocalic environment,for intrinsic

duration of the contextdid notacross

providedsee whether

inWhile

in order tosilencenaturalassociated wi thquences to date seemsHowever he examined

(io 0' clusterswas concerned

his results relshorter when the was

were also shorter \..rhen thealveolar or labial In addition,which he asvariations in vowel duration( /bVC 1 ) , the closurein the

ironments

183

results: It may be that the amount of silence listeners need in perception ismore directly related to the closure preceding the release ("C1 closure") thanto the total closure duration in production.

Method

Subjects. The subjects were two female research assistants, both nativespeakers of American English, and the author. The author, a native speaker ofthe Viennese variety of German, has lived in the United States for over 11years but has retained a foreign accent. However, it was considered unlikelythat the pronunciation of voiced stop consonant sequences in meaninglessisolated disyllables would show any systematic influence of native language.

Utterances. The utterances were the same as in Experiment 1. The 18disyllables were arranged into 10 different random lists that were typed ontoa sheet of paper in simple spelling (e.g., abdi, adgu, etc.). After listeningto sample pronunciations and practicing for a few minutes, the subjects readfrom the lists at an even pace, pronouncing each utterance at a fairly fastrate, with stress on the second syllable but without neutralizing the initialvowel. The recordings were made in a soundproof booth, using a Shuremicrophone and an Ampex AG-500 tape recorder.

Measurement procedure. All measurements were performed on a large-scaleoscillographic display provided by a GT40 computer. After inputting anutterance from audio tape, critical points in its digitized waveform werelocated in the continuous, magnified display by means of a cursor, and thedistance from one critical point to the next was measured to the nearest tenthof a millisecond using an automatic counter. Seven measurement points weredefined:

A. Approximate onset of utterance.B. Offset of VC portion. (Sometimes, vOlclng pul ses persisted into theclosure; in this case, the onset of significant damping--indicatingclosure of the vocal tract--was taken as the criterion.)C. Onset of C, release burst.D. Offset of C, release burst (approximate within a few msec) .E. Onset of CV portion.F. Onset of periodicity in CV portion.G. Approximate end of utterance.

From these measurement points, the following durations were derived:

F A Total utterance.B A::: VC portion.D B Total closure

C B::: "C, closure".DC::: C, release burst.E D "C2 closure"

G E CV portion.F E C2 burst and aspirationG - F ::: CV voiced portion.

All measurements were performed by a research assistant (astudent in s) after thorough consultation with the author.

184

graduateAnalyses

of variance were performed on all measures of interest, wi th the factorsSpeakers, Vowels (three final vowels), and Consonants (six combinations) @Since C1 and C2 were not factors, their separate influences wereexamined in post-hoc (Newman-Keuls) tests comparing those six pairs ofutterances that differed in one component only (C 1 effects: Ibgl vs@ Idg/,/bdl vs@ Igd/, Idbl vs@ ; C2 effects: Igbl vs@ Igd/, Idbl vs Idg/, Ibdlvs@ Ibg/)@ The within-cell variance (10 observations per cell, i@e.,per utterance) was taken as the error term@ Missing values, due to raremispronunciations or acoustic anomalies, were replaced wi th the cell meanprior to analysiso

and ------Total of average closure durations as a

function of and final vowels is shown in Figure 2,separately for each speaker The grand average duration was 168 msec, with anaverage within-cell standard deviation of 15 msec@ Statistical analysisrevealed, first of all, a effect, F(2,486) 262@3, p «.001: BHR'sclosures were longer (188 msec on the average) than DK's (162 msec) and SP's(154 msec) More there was a highly ficant vowel effect 9

F(2,486) = 36 1, «0001: Closure durations were shorter for final 141 (160msec) than for Iii (172 msec) and lui (1 msec) 0 This effect was shown(on the average) by all three and by each individual consonantcombination; no statistical interaction the vowel effect approachedsignificance , there was a ficant consonant effect, F(5,486) =8.5, p < .001 which did not interact lrli th any other factor, despite (orbecause the considerable variabil evident in Figure 2. The sixconsonant combinations were as follows: (161 msec), (165msec) , Idbl (1 (172 msec) , Ibdl (173 msec)@Newman-Keuls effect of the first (/dlshorter than Ibl ficant effects of the second stop (/glshorter than Ibl of three sons in eachcase.

closures to be shorter inor to be especially short

, the results are in excellent(Note 1) measurements, which showed closures to be

in first and for velar stops in secondfound, in with results, that

in context, and he related this findingdurations We will return to

, these databack-front sequences than inin l~dbo.l and (cf

with'sshortest for alveolar

alsowere

half ofbursts,

BHRTheir

standardin BHR'

001,above.

185

BHR

- OK()(1)(/)

E

z0

cr::::J0

6-6/UI

0---01 if011111111110 fu f

combinations producedtotal closure durations in 18 VC 1

186

al though there had been such an effect on total closure duration However,there was a highly significant effect of consonants, F( 5,324) :: 3407, .E. <0001, which also interacted with t F(1,324) :: 307, p < 01. Averagingover speakers (which seems permissible since the interaction was quite small),the rank order was: Igbl (63 msec) , Idbl (66 msec) , Idgl (70 msec) , Igdl (72msec) , Ibgl (80 msec) , Ibdl ( msec)., Newman-Keuls tests showed C1 closuresto be clearly longer when C1 was Ibl than when it was Idl or Igl (p < .01)--aresult that is in striking with measurements of closure durations insingle intervocalic stops, which show longer durations for labials (e.g G ,

Kohler, 1979; Umeda, 1977; Westbury, Note 1) 0 However, C2 also affected C1closure duration: C1 closures were longer preceding Idl than preceding Ibl (p< 001) or Igl (p < .01, but only shown DK). Thus, while asyllable-final Ibl led to long C1 closures, a following syllable-initial Iblwas associated with rather short C1 closure durations@

.£2 closure" The C2 closure measurements for speakers BHR and DK areshown In the right half of Figure 30 The average C2 closure lasted 84 msec,with an average within-cell standard deviation of 16 msec" BHR's C2 closureswere significantly longer (90 msec) than DK's (78 msec) , F(1,324) :: 46.1, £ «•001, as had been his C1 closures.. There was a ficant vowel effect,F(2,324) 609, p < .001, C2 closures being shorter preceding 10.1 (80 msec)than preceding li7 (85 msec) or lui (87 msec) Since C1 closure had shown novowel effect, it was C2 closure that was responsible for the variations intotal closure duration with final vowel.. C2 closure durations varied signifi­cantly across different consonant combinations, F(S,324) :: 13 .. 2, p < 0001, andthe pattern differed somewhat between the two speakers, F( 5,324):: 4.2, £ <.001. Overall, however, the rank order was nearly the inverse of that for C1closure duration., Ibdl (75 , Idgl (78 msec) , Ibgl (82 msec) , Igdl (84msec) , Idbl (90 ,/gbl ( Newman-Keuls tests showed thatsyllable-initial Ibl (C2) was associated with longer closures than eitherIdl or Igl (p < .. 01), with somewhat closures for Igl than for Idl (p <.05) , - closures were shorter the ing stop was Ibl thanwhen it was < .01) Thus closures, like C closures, were longestwhen the was lab ,but tended to short when the otherstop was labial.

closure duration measures are directlythe other measurements will be summarized

(average duration 17 msec) were markedly~~------'s utterances, but not in DK' s. VC

showed no difference (in contrastof C1 The vocalic portion was shorter

.01) The andthe effect of( 11 and for

Two (BHR and had

187

j-l

0000

c c SURE C2 C SURE

BHR

0 .. '''' ....0 lui

6.-6./0/

o---o/il

gd

A

:c.

gbdbdgbg

o~

"' ...,0.,"-~ ..0 """ 0 ..

....6. ..,. ".., '''.p",

, 0

""".0....a.......'~".. ~,,,'r' OK',....,r:>/......," I D.~A"..

......,'" . .. ..0' ..~ ....

...... .6>.____ "0

..~:1 ----~ ,IL~_-·6 ~

0 ....'....0"

OK

BHR

gd bd

UTTERANCES

gbdbdgbgbd

()(j)(/'J

Z0I-«0:::J0

70

60

e 30 Average C, and C2 closure durations in 18 VC 1-C2V combinationsproduced by two speakerso

Summary. Closure durations were affected by the identity of bothconsonants as well as by the final vowel. C1 and C2 generally had oppositeeffects; thus, total closure durations ranked Idl < Ibl < Igl with respect toC1 but Igl < Ibl < Idl with to C2. When C1 and C2 closure segmentswere considered , however the consonant effects were found toreflect primarily labial articulation: Both C1 and C2 closures were longestwhen the associated consonant was Ibl and tended to be shortened when theother was Ib/" Total closure durations were shortest in I -oJ context,and this effect was entirely due to variations in C2 closure.

This pattern of results does not show a close resemblance to theresul ts of riment 1. The abnormal perceptual boundaries for

ladbaJ and IQ have no in , and systematic effects of C1and across all three vocalic contexts are observed in production only, notin perception" Only the final-vowel effect (shorter closures in I-al context)corresponds to a tendency towards shorter perceptual boundaries in thatcontext. However, this effect could have an auditory basis: Severalstudies have shown that silent gaps are easier to detect in spectrallyhomogeneous than in environments (Collyer, 1974; Perrott &Williams, 1971; Williams & Perrott, 1972) Since the initial vowel in the

stimuli was stimuli in were spectrally morehomogeneous than stimuli in I-il or l-u/, and perhaps this homogeneityfacilitated the detection of the silent closure period.

EXPERIMENT STIMULI---So • 1) and production (Exp. 2) of

However, the results of Experimentdue to ari ties of the syntheticseems in view of the good

resul ts of 1 and the earlierand Dorman (in press), it seemed

stimuli. This was

They included ten paid vo­experiments and two

listeners (a research

Twelvelittle

with considerableassistant and the author)

theclosure

various amounts ofVC and CV

BHR's utterances,2 To avoid token-

natural variabil ,fourwere selected from the 10

consisted 4 x 18kHz and ed i ted

The, and

189

The experimental tapes were analogous to those of Experiment 1. Threeparallel sets were recorded, one for each final vowel. In each set, the firststimul us sequence consisted of the isolated VC and CV portions in randomorder, arranged in 5 blocks of 48, wi th ISIs of 2.5 sec and 10 sec betweenblocks. The 48 stimuli resulted from 4 tokens of each of 2 portions (VC andCV) of 6 utterances. The second stimulus sequence contained the VC-CVcombinations in random order, arranged in 4 blocks of 66, with ISIs of 2.5 secand 10 sec between blocks. The 66 stimuli resulted from 11 closure durationsfor one token of each 0 f the sit{ utterances. Di fferent tokens were used ineach of the four blocks; thus, there were in fact 4 x 66 :: 264 physicallydifferent stimuli.

Procedure. Each sUbj ect pa rtic ipa ted in three sessions, one for eachfinal-vowel condi tion. The order of final-vowel condi tions was counterbal­anced across subjects. In each session, the isolated VC and CV portions werepresented first. A total of 5 responses for each token of each 'utterance wasobtained, i.e., 20 responses for each utterance when token variation isignored. Sub sequently, the VC-CV combinations were presented three times,separated by appropriate rest periods. That is, each subject gave a total ofthree responses to each individual stimulus, or 12 responses when ignoringtoken variation.

Results and Discussion

Monosyllables. The natural-speech CV stimuli were qui te intelligible,but the VC stimuli were less well identified than the synthetic stimuli inExperiment 1 The stop in IQg/, in particular, was frequently misidentified,wi th "b" confusions being about twice as frequent as "d" confusions. Thispoor identifiability was obviously a consequence of removing the C1 releaseburst. The percentages of correct responses for the ICi/, I Ca.l , ICu/, andlaCI sets were 90.2, 96.3, 99.4, and 82.3, respectively (52.1 for I~g/). Theconfusion patterns did not seem to reflect in any way the context in which agiven stimulus portion had been pronounced; thus, there seemed to be littlecoarticulation between VC and CV portions.

VC-CV combinations: Two-stop vs. one-stop responses. The results of themain part of the experiment were somewhat startling. Although the twoexperienced subjects produced what seemed to be typical and orderly results, anumber of the naive subjects failed to show the VC-CV interference phenomenon,i oe 0' the predominaQce of single-stop percepts at short closure durations.All naive subjects reported two stops at short silence durations for at leastsome of the stimuli Moreover, these responses were correct more often thannot, and those misperceptions that occurred were typically consistent andstimulus-specific

This outcome was quite unexpected, even though it will be recalled thattwo subjects in Experiment 1 had to be excluded for the same reason. To makesure that no problem of instructions was involved, two of the subjects wererecalled and carefully instructed by the author. The same result wasobtained· There were very few single-stop responses. of thestimuli did not reveal any reason for this "abnormal" behavior of the majoriof listeners. Of course, researchers have known for a long time that speechcues--silence in particular--do not always have a perceptual effect: Their

190

effect depends on the values of other relevant cues in the signal a In thecase. the formant transitions in and out of the closure and the C2

release burst may have provided manner cues strong enough to override theperceptual effect of silencea What is surprising is that this occurred onlyfor the naive listeners. as if less to the silence cuethan the two listeners e • very similar observationshave Porter. and Miller (1980)a

either gave no single-stoprandomly distributed a

five naive listeners fell into amoreover resembled the resul ts of the two

Therefore, the data of all seven sUbjects4, which is analogous to Figure 1.

Five ects had to be excluded becauseresponses at all or just a few thatHowever \l the responses of the

thatlisteners, BHR and

were combineda are

the naiveconsistentschedulewas still

The first shows a vowel effect, F(2,12) = 4.7. ~ < .05:Considerably more silence was to hear both stop consonants in I-ilcontext than in and I-ul contexts, and sl y less silence was requiredin context than in I-ul contexte While the latter tendency parallels thefind f 1 and the first difference has no

in the earlier results This difference was primarily due toects since neither BHR nor PP showed any vowel effects that were

across all six consonant combinations a Inspection of the testthat the effect was not an artifact of test order, which

balanced across the ected ects

The secondsix vc-cvacross the threes ficant e) Inranked second indurations were in Ibdlhad the shortest boundary. Oncefollow the front-back v back-frontan effect of C2 : silenceswas either Ibl or < .01 inrank order I > Ibl with robtained in production, indclosure durations inruns counter to

of differences across the< .001, that te consi stent

interaction was marginallywere red for ; Ibgl

third. The shortest silencein I-ul context where Idb/

does not consistentlynction. Rather, it seems to reflect

when was /g/ than when itthat the boundary

is preci the opposite of thatcombinations with totalsilence in This

conceived at the outset

191

!-'\.0N

- -

0---0 ilOnuuuuO lui

gd

LI

4 Percent single-stop responses (averaged over all silence durations)to the 18 VC1 combinations (natural speech).

misidentified. What isat all silence durations,

in the direction ofthe listeners seemed to "know"

into a singleC1 so inexperiment resembled

course, the perception ofnoteworthy is that a 1includ the , and thattwo stops, rather than one" In other words,that confl VC and CV cues could not beit is not clear, however, what led them toVCCV context Note that the ttern inthat found in 1.

red to hear two stopsto be ated with

natural utterances. Theyemploying synthetic and

the cause for the perceptual variabilitystimuli; it does not seem to be

s Presumably, theeffective value of some other

the acoustic environment (indecision process.

variations in theof the VC,

closure durdeal between

in utterancesvariations ineven differ anatural stimuli,must be sought in aud

ed in listeners Y

effective amount of silencerelevant stimulus characteristic,ways not und before it

Thiaudi toryA number

, we1 willbetween these twomost e andresearch

forti and lenis

and monoticconstraints?

No e 1) i